Re: [External] Re: [PATCH v3 1/6] mm: migrate: do not migrate HugeTLB page whose refcount is one

[Muchun Song <songmuchun@bytedance.com>]

On Tue, Jan 12, 2021 at 10:28 PM David Hildenbrand <david@redhat.com> wrote:
>
> On 12.01.21 15:17, Muchun Song wrote:
> > On Tue, Jan 12, 2021 at 9:51 PM David Hildenbrand <david@redhat.com> wrote:
> >>
> >> On 12.01.21 14:40, Muchun Song wrote:
> >>> On Tue, Jan 12, 2021 at 7:11 PM David Hildenbrand <david@redhat.com> wrote:
> >>>>
> >>>> On 12.01.21 12:00, David Hildenbrand wrote:
> >>>>> On 10.01.21 13:40, Muchun Song wrote:
> >>>>>> If the refcount is one when it is migrated, it means that the page
> >>>>>> was freed from under us. So we are done and do not need to migrate.
> >>>>>>
> >>>>>> This optimization is consistent with the regular pages, just like
> >>>>>> unmap_and_move() does.
> >>>>>>
> >>>>>> Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> >>>>>> Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
> >>>>>> Acked-by: Yang Shi <shy828301@gmail.com>
> >>>>>> ---
> >>>>>>  mm/migrate.c | 6 ++++++
> >>>>>>  1 file changed, 6 insertions(+)
> >>>>>>
> >>>>>> diff --git a/mm/migrate.c b/mm/migrate.c
> >>>>>> index 4385f2fb5d18..a6631c4eb6a6 100644
> >>>>>> --- a/mm/migrate.c
> >>>>>> +++ b/mm/migrate.c
> >>>>>> @@ -1279,6 +1279,12 @@ static int unmap_and_move_huge_page(new_page_t get_new_page,
> >>>>>>              return -ENOSYS;
> >>>>>>      }
> >>>>>>
> >>>>>> +    if (page_count(hpage) == 1) {
> >>>>>> +            /* page was freed from under us. So we are done. */
> >>>>>> +            putback_active_hugepage(hpage);
> >>>>>> +            return MIGRATEPAGE_SUCCESS;
> >>>>>> +    }
> >>>>>> +
> >>>>>>      new_hpage = get_new_page(hpage, private);
> >>>>>>      if (!new_hpage)
> >>>>>>              return -ENOMEM;
> >>>>>>
> >>>>>
> >>>>> Question: What if called via alloc_contig_range() where we even want to
> >>>>> "migrate" free pages, meaning, relocate it?
> >>>>>
> >>>>
> >>>> To be more precise:
> >>>>
> >>>> a) We don't have dissolve_free_huge_pages() calls on the
> >>>> alloc_contig_range() path. So we *need* migration IIUC.
> >>>
> >>> Without this patch, if you want to migrate a HUgeTLB page,
> >>> the page is freed to the hugepage pool. With this patch,
> >>> the page is also freed to the hugepage pool.
> >>> I didn't see any different. I am missing something?
> >>
> >> I am definitely not an expert on hugetlb pools, that's why I am asking.
> >>
> >> Isn't it, that with your code, no new page is allocated - so
> >> dissolve_free_huge_pages() might just refuse to dissolve due to
> >> reservations, bailing out, no?
> >
> > Without this patch, the new page can be allocated from the
> > hugepage pool. The dissolve_free_huge_pages() also
> > can refuse to dissolve due to reservations. Right?
>
> Oh, you mean the migration target might be coming from the pool? I guess
> yes, looking at alloc_migration_target()->alloc_huge_page_nodemask().

Yeah, you are right. If we want to free a HugeTLB page to
the buddy allocator, we should dissolve_free_huge_page()
to do that. Migrating cannot guarantee this at least now.

>
> In that case, yes, I think we run into a similar issue already.
>
> Instead of trying to allocate new huge pages in
> dissolve_free_huge_pages() to "relocate free pages", we bail out.
>
> This all feels kind of wrong. After we migrated a huge page we should
> free it back to the buddy, so most of our machinery just keeps working
> without caring about free huge pages.
>
>
> I can see how your patch will not change the current (IMHO broken) behavior.
>
> --
> Thanks,
>
> David / dhildenb
>